Efficiently estimating the protein nitrogen content of rice leaves (LPN) is crucial for monitoring the nutritional health of rice and guiding precision fertilization based on requirements. Unmanned aerial vehicle (UAV)-acquired hyperspectral imagery is a key tool for estimating rice nitrogen content. Previous studies have demonstrated the potential of machine learning models for this task. However, these models typically require substantial data for supervised training to ensure high performance and generalizability. Acquiring a large sample size is challenging due to weather conditions, high collection costs, and other factors. Moreover, machine learning models have low interpretability. Enhancing it is vital for understanding the model’s decision-making. To address these issues, we utilized the Wasserstein-generative adversarial network (WGAN) algorithm to expand the sample dataset. This method employs statistical regression (multiple linear regression (MLR) and partial least squares regression (PLSR)) and machine learning (support vector machines (SVM) and K-nearest neighbor (KNN)) algorithms to establish an estimation model for the LPN. The Shapley Additive exPlanations (SHAP) method was used to analyze the contributions of the input features to LPN estimation. An experiment was conducted at the National Agricultural Science and Technology Park, Guangzhou, Baiyun District, Guangdong, China. The model based on the KNN provided the optimum estimation performance, and the model accuracy was improved by adding the augmented dataset, resulting in a 10.39% improvement in the R 2 value. The SHAP values revealed that B 775.6 , double-peak canopy nitrogen index (DCNI), and MERIS terrestrial chlorophyll index (MTCI) were the core variables for LPN estimation. These findings provide significant references for precision fertilization and improving nitrogen use efficiency in rice cultivation.
Peng et al. (2026) studied this question.